JPH05346855A - Inference device - Google Patents

Abstract

PURPOSE:To shorten time for collating patterns by effectively using the result of pattern collation. CONSTITUTION:When a rule selecting means 13 selects a rule having high possibility to be started among respective rules stored in a knowledge base 12, a pattern collating means 14 performs pattern collation between the condition part of this selected rule and data. The collated result of this pattern collation is stored in memories 16 and 17. Based on data stored in these memories 16 and 17, a collating condition managing means 15 manages and corrects data to be changed by rule execution among data stored in the memories 16 and 17 after the operation part of the rule is executed, and information showing how much the pattern collation is progressed concerning respective rules is recorded and updated as the collating condition. Then, the cycle of recognition- execution is repeated until no rule in which the condition part of rule and data are matched is in existence.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an inference device which executes inference based on a rule consisting of a condition section and an operation section.

[0002]

2. Description of the Related Art Rule-based reasoning is represented by a cognition-execution cycle as shown in FIG. 7, which is applied to a human cognition / interpretation / action cycle. FIG. 8 schematically shows a recognition-execution cycle of such an inference device. That is, the inference device has a working memory 1 and a knowledge base 2.
Is provided. Of these, the working memory 1 stores, for example, respective data x1 and x2 such as temperature and pressure in the tank.
Is stored in the knowledge base 2. For example, "x1> A → W" is stored in the knowledge base 2.
Each rule such as "(temperature x1 is A degree or more, water supply W)" is stored. In this case, x1> A is the condition, and W is the operation unit.

In the inference control in the state where each data is stored, pattern matching is performed with each rule for each data x1 and x2 stored in the working memory 1. This pattern matching is performed between each data x1, x2 and the condition part of each rule. As a result of this matching, each rule "x1> A → W""x2 = C → Y"

Is selected. Here, there are two selected rules, and a set of conflict rules is formed. Next, the conflict of these rules is resolved. For this conflict resolution, for example, a rule with a higher priority of each rule is selected, and for example, the rule "x1> A → W" is selected. Then, the operation unit W of this rule is executed, and one cycle of the recognition-execution cycle ends. Since the data stored in the working memory 1 changes due to the execution of this rule, the pattern matching between each data and each rule is performed again, and the recognition-execution cycle is repeated until there are no executable rules.

By the way, there is a RETE algorithm as an algorithm for performing such inference at high speed. The RETE algorithm saves the result of pattern matching performed once and avoids performing the same pattern matching again, thereby performing inference at high speed. In other words, at the time of pattern matching for selecting one rule to be executed, all the pattern matching results of each rule and all data are stored,
If the data in the working memory 1 has changed, all patterns are re-matched only for the changed amount. From the set of conflict rules obtained from the result of this pattern matching,
For example, a rule to be executed is selected according to the priority. When the operation section of this rule is executed and the data in the working memory 1 changes, pattern matching is performed,
The cognitive-execution cycle repeats until there are no viable rules.

Therefore, in the above-described inference control, every time one rule recognition-execution cycle is performed, pattern matching is performed between the rule and all the changed data to form a complete set of conflict rules. Therefore, pattern matching takes a very long time.

Further, since all the pattern matching results are stored, when the data in the working memory 1 changes due to the rule execution, the pattern matching results performed up to that point are deleted before the rule is executed. Can be wasted.

[0008]

As described above, pattern matching takes a very long time, and since all pattern matching results are stored, the pattern matching results may be deleted and wasted. Therefore, an object of the present invention is to provide an inference device that can effectively use the result of pattern matching to shorten the time of pattern matching and reduce the memory capacity.

[0009]

SUMMARY OF THE INVENTION The present invention comprises a knowledge base for storing each rule and a working memory for storing each data, and a rule having a matched pattern after pattern matching between the condition part of the rule and the data is performed. In an inference device that repeats the cognitive-execution cycle for executing the operation unit of

A rule selecting means for selecting a rule that is likely to be activated among the respective rules, a pattern matching means for performing pattern matching between the condition part of the rule selected by this rule selecting means and the data, and this pattern. A memory for storing the result of matching by the matching means, and a matching status managing means for storing and managing as a matching status how far the matching of each rule has progressed after the rule operation section is executed based on the data stored in this memory. It is an inference device which is equipped with and tries to achieve the above object.

[0011]

With the provision of such means, when the rule selecting means selects one of the rules stored in the knowledge base that is likely to be activated, the condition part of this selected rule is selected. The pattern matching means performs pattern matching with the data. The matching result of this pattern matching is stored in the memory, and after the operation part of the rule is executed based on the data stored in this memory, the changing data among the data stored in the memory is corrected and The matching status managing means corrects and manages the matching status of how far the pattern matching for the rule has progressed. Then, the recognition-execution cycle is repeated until there is no rule in which the condition part of the rule and the data match.

[0012]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of an inference device. The inference engine 10 has a working memory 11 and a knowledge base 1.
2 is connected. Each data is stored in the working memory 11. Each data stored in this working memory 11 is (class attribute ...)

Is represented by Here, the class represents the type of data such as human or dog, and the attribute (hereinafter, at
(represented by r) is the attribute value of the class, such as name or age,
It represents your height. Hereinafter, when each data [0] to [5] is represented by (class atr1, atr2), [0] (A2) [1] (A-10) [2] (A6) [3] (B 1 3) [4] (B 1 4) [5] (B 1 -1). For example, data [0] represents class “A” and atr1 represents “2”, and data [3] represents class “B” in which atr1 represents “1” and atr2 represents “3”. On the other hand, the knowledge base 12 stores the following rules.

[0014]

[Equation 1]

Here, in the rule, the front side up to "→" represents the condition section, and the rear side represents the operation section. For example, the meaning of the rule "1" is that the class A in the working memory 11 is at
If there is one with r equal to or greater than “0”, and if atr1 is “1” in class B and atr2 is larger than the value x of atr in class A, the working memory of the target class A 11 This is the content of deleting the data of class A in.

Further, in the numerical value described after the rule name, for example, "50" in the rule "1" indicates the priority of the rule, and the larger this value is, the higher the priority is.

The inference engine 10 has a function of repeating the recognition-execution cycle of executing the operation part of the rule-matching pattern after performing pattern matching between the condition part of the rule and the data until there is no matching rule. In addition to this function, the inference engine 10 has the functions of the rule selection means 13 and the matching status management means 15.
Further, an α memory 16 and a β memory 17 are connected to the inference engine 10. The rule selection means 13
Of the rules stored in the knowledge base 12,
For example, it has a function of selecting rules according to the order of priority.

The pattern matching means 14 has a function of performing pattern matching between the condition part of the rule selected by the rule selecting means 13 and the data. Then, the data matching the condition part of the rule obtained during the matching by the pattern matching means 14 and the data matching the condition part of the rule obtained as a result of the matching are stored in the α memory 16 or the β memory 17. It has become. In addition,
The α memory 16 is dedicated to storing individual data, and the β memory 17 is dedicated to storing each data combination.

Further, the collation status management means 15 records and updates the pattern collation status indicating to what extent (for which data and under what condition) the pattern collation for each rule has been performed, and the result is added to the data. It has a function to correct the collation result and collation status based on the operation. Next, the inference control of the device configured as described above will be described.

In this inference control, each rule above "1"
I will explain the reasoning by "2", but these rules "1"
The inference control at "2" is executed according to the development view of the RETE network shown in FIG.

By the way, in the figure, “◯” is an intra node indicating a comparison item related to only one data stored in the working memory 11. For example, “Is the class A?” “The value of atr is 0 or more. "?" Is made. “⊚” is an internode indicating a comparison item that must be satisfied by different data. For example, a comparison is made such as “is the value of atr2 of class B greater than the value x of atr of class A?”. .. Further, each node indicated by “Δ” and “▲” is for storing in the α memory 16 and the β memory 17, respectively, and the intra-node and the inter-node existing on the path to these nodes are matched by the pattern matching. , The data arriving at that node is stored. The node indicated by “□” is the end of the network and corresponds to each rule “1” and “2”. If there is data that has reached this node "□", the rule "1" or "2" can be executed. The inference control will be described below with reference to the flow chart of the inference control shown in FIG.

In step # 1, the rule selection means 13
Refers to the pattern matching status in order to exclude the pattern matching rule, but since the pattern matching has not yet been performed in the first recognition-execution cycle, the knowledge as a rule that is likely to be executed next is Base 12
The rule 1 having the highest priority is selected from the respective rules stored in.

Next, proceeding to step # 2, the pattern matching means 14 first sets each data [0] to
Pattern matching is performed to determine whether the class is A with [5], and the data [0] to [2] are sent to the next node. Next, in the node N1, pattern matching is performed between the condition of class A described in the condition part of the rule “1” and each data [0] to [2]. As a result of this pattern matching, "the class is A
So is the value of art 0 or more? The data satisfying the condition of “” is [0] [2], and the data not satisfying the condition is [1]. As a result, the pattern matching means 14 stores each data [0] [2] in the area M1 of the α memory 16 as shown in FIG.

Next, the pattern matching means 14 determines the node NB.
In, the data [0] to [5] is compared with the pattern whether the class is B or not, and the data [3] to [5] are sent to the next node. In node N2, the condition of class B described in the condition part of rule "1" and each data [3]
~ Pattern matching with [5] is performed. As a result of this pattern matching, the data satisfying the condition of "class is B and atr1 is 1" is [3] [4] [5], and accordingly, the pattern matching means 14 sets each data [3] [4]. [5] is stored in the area M2 of the α memory 16.

Next, the pattern matching means 14 uses the α memory 16
Data [0] [2] stored in the respective areas M1 and M2 of
And [3], [4], and [5] at the intranode N4, that is, each condition “class is A, is art value 0 or more?” And “class is B, and atr 1 is 1, atr. Two
The value of is greater than the value x of the atr of class A? Pattern matching is performed to select a combination of data that both satisfy ". As a result of this pattern matching, data satisfying both conditions are a combination of ([0] [4]) and ([0] [3]).
It becomes two sets of the combination of. Therefore, the pattern matching means 14 stores the combination of these data in the area M3 as shown in FIG.

Thereafter, the inference engine 10 executes the rule "1" in step # 3. In this case, whichever of the data combinations ([0] [4]) and ([0] [3]) may be executed first, for example, the previously obtained data ([0] [4] ]) Is executed.

Next, the process proceeds to step # 4, inference engine 1
The matching status management unit 15 of 0 stores the completion of the pattern matching for rule 1 in the matching status memory, and each data [3] according to the data flow for rule “1”.
[4] It is confirmed that [5] has flowed to the node N2 and is stored in the collation status memory, and at the same time, that the collation of the node N3 and below relating to the rule "2" is not completed and the situation is collated. Save to situation memory. Since the execution result data [0] is deleted, the combination data including [0] in M1 and M3 is deleted. As a result, there is no combination that can be executed for rule “1”, and the fact that execution has ended is saved. As described above, one cycle of the rule "1" is completed, and the process returns to step # 1.

Next, the rule selecting means 13 knows that the matching and execution of the rule "1" have already been completed depending on the matching status, so that the rule which is likely to be executed next is
From the rules stored in the knowledge base 12, the rule "2" having the next highest priority is selected. Hereinafter, the above steps # 1 to # 4 are executed, and the recognition-execution cycle is repeated until there is no matching rule.

As described above, in the above embodiment, the rule selecting means 13 selects a rule having a high priority among the rules stored in the knowledge base 12, and the pattern matching between the condition part of this rule and the data is performed. Is performed by the pattern matching means 14, and the result of this matching is stored in each memory 17,
Since the rule operation unit is executed based on these data stored in the memory 18, the changing data among the data stored in the memories 17 and 18 is corrected and managed by the collation status management means. The range to be subjected to the pattern matching is limited, and the pattern matching is performed only in this range, so that the processing amount of the pattern processing required to execute one rule can be reduced.
Further, it is possible to reduce unnecessary pattern matching due to correction of pattern matching performed halfway and change of data in the working memory 11. Therefore, the capacity of the α and β memories 17 and 18 required at the time of execution can be reduced.

The present invention is not limited to the above-mentioned one embodiment, and may be modified within the scope of the invention. For example, rule conflict resolution is not limited to priority, but a method of preferentially selecting a recently executed rule or a method of selecting a rule similar to the condition part or operation part of a previously executed rule may be used. Good.

It is also possible to use the certainty factor of rules and the order of description according to the method for resolving rule conflicts. For example, if the time tag or certainty factor of working memory data is used for the rule conflict resolution method, check the time tag in each class of working memory data, and check the working memory data described in the condition part of the rule. Then, pattern matching may be performed from a rule that has many new ones.

Further, the pattern matching range may be limited by selecting or ordering the data based on the conflict resolution method. For example, if the rule that matches the latest data is given priority, the data is sorted in order of time tags and pattern matching is performed in order. If even one executable rule is obtained, pattern matching is stopped at that point. You may do it.

[0033]

As described in detail above, according to the present invention, it is possible to provide an inference apparatus that can effectively use the result of pattern matching to shorten the time of pattern matching and reduce the memory capacity.

[Brief description of drawings]

FIG. 1 is a configuration diagram showing an embodiment of an inference device according to the present invention.

FIG. 2 is a development view of a Rete network of rules executed by the device.

FIG. 3 is a flow chart of inference control in the device.

FIG. 4 is a diagram showing pattern matching processing in the apparatus.

FIG. 5 is a diagram showing a pattern matching process for one rule in the apparatus.

FIG. 6 is a development view of a normal RETE network executed by a conventional device.

FIG. 7 is a diagram showing a recognition-execution cycle of inference control.

FIG. 8 is a schematic diagram of inference control.

[Explanation of symbols]

10 ... inference engine, 11 ... working memory, 12 ...
Knowledge base, 13 ... Rule selection means, 14 ... Pattern matching means, 15 ... Matching status management means, 16 ... α memory, 1
7 ... β memory.

Claims (1)

[Claims] 1. A recognition system comprising a knowledge base for storing each rule and a working memory for storing each data, and performing a pattern matching between the condition part of the rule and the data, and then executing an operation part of the rule having a matched pattern. In an inference device that repeats execution cycles until there are no more matching rules, rule selection means for selecting a rule that is more likely to be activated among the above rules, and condition parts and data for the rules selected by this rule selection means Pattern matching means for performing pattern matching with the pattern matching means, a memory for storing the matching result by the pattern matching means, and the rule operation section based on the data stored in this memory Inference, characterized by having a collation status management means for storing and managing the progress as a collation status Location.